Classification of wolf call types using remote sensor technology

A system is presented for the near‐real‐time classification of calls from captive Alaskan (Canis lupus occidentalis) and Mexican (Canis lupus baileyi) wolves as well as other ambient sounds such as bird calls and anthropomorphic noise. Signals are detected on a Viper network node at The California Wolf Center using a signal‐to‐noise ratio based call activity detector. These calls are transmitted over a regional high‐speed wireless network (HPWREN) to a remote processing facility located 50 miles from the collection site. Cepstral feature vectors are extracted from candidate calls at the remote facility and are classified using a hidden Markov model pattern recognition system trained with manually labeled data representing the wolf ethogram and other common noises in the soundscape. In an offline evaluation with known segmented data, the system achieves an 84% accuracy. [This work is supported by HPWREN, NSF Grant No. 0426879.]